Data and AI releases

This month963Companies169

963 this month · 169 companies

truss 0.18.34

UpdatetodayUnverifiedAdded Oct 13, 2026

- Add include and exclude filters to BDN mount configuration feat: Forward egress_restrictions for BIS and warn about missing hosts Release 0.18.34

Baseten's release notes

More truss releases

Every truss release
Updateunverified

truss 0.18.33

- feat(performance-client): NET-1589 Explicit trace context and opt-in OTLP export fix(performance-client): stop hedge and retry budgets from wrapping past zero Add --spot to truss train workstation fix(chains): consolidate push logging and stop duplicate status header Add an SSH opt-out for loops exec Keep inference working when performance client tracing...

Updateunverified

truss 0.18.32

- Bump control server truss pin to fix dev deployments on newer GPUs Set control server websocket limit to 100MiB (NET-1566) Fix two integration test failures on main Allow data-parallel BEI encoder deployments Release 0.18.32

Updateunverified

truss 0.18.31

- Truss: support updated training auth config (OIDC, AssumeRole) Remove transfer command from BEI startup trt config: Honor configured BEI max token limit Add fp8_mlp_only quantization type fix(config): preserve YAML merge keys during duplicate checks Honor global --non-interactive in child commands Update the base_image examples in the config schema Add...

Updateunverified

truss 0.18.30

- fix(trt-llm): Add LoRA cache runtime settings Use relative references for BDN volume mounts fix(performance-client): allow opting out of global tracing initialization Add BDN access and hotload configuration under BND section Release 0.18.30

Updateunverified

truss 0.18.29

- feat(cli): use CUDA 12.9 as the workstation default feat(cli): add JSON output to truss train workstation Release 0.18.29

Updateunverified

truss 0.18.28

- Tolerate an empty teams list during team resolution Keep external_data Dockerfile dest as a POSIX container path Add bnd section to Truss config Align README quickstart with the TRT-LLM push path Release 0.18.28

Updateunverified

truss 0.18.27

- Isolate docker server dependencies from base image uv config feat(cli): default B300/GB300 workstations to CUDA 13 image Speed up ignored path traversal in watch mode Fix shell injection in external_data Dockerfile curl RUN lines Add AWS_ASSUME_ROLE auth method for base images and weights Release 0.18.27

Updateunverified

truss 0.18.26

- feat(cli/train): show capacity type (spot/on-demand) in train view fix(bis): Push model_metadata for BIS trusses perf client fix feat(performance-client): support msgpack bin in untyped responses Warn only once per process when keyring write fails Revert "Add request backpressure policy to truss" Support authenticating truss via env vars as an alternative...

Also shipped on Oct 13, 2026

Baseten in October 2026
Betaunverified

What's included:

- Three training methods: Supervised fine-tuning (SFT) , direct preference optimization (DPO) , and reinforcement learning with group relative policy optimization (GRPO) . Each method trains a low-rank adapter (LoRA) or the full model weights. Custom loss functions : Train with any objective you compute in PyTorch from the model's log-probabilities.

Updateunverified

LangSmith SDK 0.14.8

- fix(py): send session_id on RunTree.patch fix: keep punctuation out of freeze guard links chore(deps): bump source-map-js from 1.2.1 to 1.2.2 in /js fix(py): make the /runs/multipart 404 fallback to /runs/batch reachable feat(sandbox): accept any JOSE asymmetric algorithm chosen by the JWKS key (python) feat(sandbox): customer-managed KMS keys for S3...

Updateunverified

Retry jobs with narrower API token permissions

Retry automation no longer needs the broader write_builds REST API scope. The new retry_builds scope allows a token to retry failed jobs in existing builds or individual eligible jobs, without granting access to the create-build or rebuild endpoints. When creating or editing an API access token, select Build Retries ( retry_builds ).

Updateunverified

Post-training

Post-training lets you write your own training loop and run it against Together AI GPUs. Your code builds each batch and chooses the loss function and learning rate, while the GPUs you provision run the forward and backward passes, optimizer steps, and sampling. It bills per GPU-minute.